Feature-Enhanced Fusion and Adversarial Domain Adaptation for Cross-Condition Few-Shot Fault Diagnosis of Bearings

Rolling bearings are core components of rotating machinery, and their operational status is critical to the safe and reliable service of major equipment. However, the variable rotational speeds and loads in industrial environments cause significant data distribution drift; coupled with an extreme scarcity of labelled fault data under new operating conditions, this leads to a severe decline in the generalization performance of traditional deep learning models based on the assumption of independent and identically distributed (i.i.d.) data. To address the dual challenges, this paper proposes a fault diagnosis method based on Feature Enhancement Fusion and Adversarial Domain Adaptation (FEF-ADA). Firstly, a time-frequency spectrum is generated via the Continuous Wavelet Transform, and high-dimensional, discriminative fault features are deeply extracted whilst suppressing industrial background noise through a combination of hollow convolution, spectral convolution and a channel-spatial dual attention mechanism; secondly, a domain adversarial network with a gradient reversal layer is introduced to align the feature manifolds of the source and target domains via a maximum-minimum game mechanism, thereby extracting operating-condition-invariant representations; finally, prototype learning is introduced in the metric space to establish accurate classification decisions relying solely on a very small number of target-domain support samples. Experiments demonstrate that, under conditions of only 15 samples per class (15-shot), FEF-ADA achieved an average recognition rate of 89.70%, with a standard deviation of just 0.70% across five independent experiments. Its performance significantly outperforms traditional CNNs (73.41%), ProtoNet (78.66%) and DANN (82.47%), demonstrating excellent resistance to operating condition drift and significant engineering practical value.